def prepare_data():
# prepare data
training_data = list(read(dataset='training', path='./'))
testing_data = list(read(dataset='testing', path='./'))
split = len(training_data)
label, pixels = training_data[0]
train_data = np.zeros(shape=(len(training_data), 1) + pixels.shape)
train_label = np.zeros(shape=(len(training_data), 10))
for n in range(len(training_data)):
train_label[n, training_data[n][0]] = 1
train_data[n, 0, :, :] = training_data[n][1] / 255.0
Te_label, Te_pixels = testing_data[0]
test_data = np.zeros(shape=(len(testing_data), 1) + Te_pixels.shape)
test_label = np.zeros(shape=(len(testing_data), 10))
for n in range(len(testing_data)):
test_label[n, testing_data[n][0]] = 1
test_data[n, 0, :, :] = testing_data[n][1] / 255.0
# Downsample training data
n_train_samples = 30000
train_idxs = np.random.random_integers(0, split - 1, n_train_samples)
train_data = train_data[train_idxs, ...]
train_label = train_label[train_idxs, ...]
return train_data, train_label, test_data, test_label
def main():
lb, im = mnist.read(dataset="training")
lbt, imt = mnist.read(dataset="testing")
im = im.reshape((len(im),len(im[0])*len(im[0])))
imt = imt.reshape((len(imt),len(imt[0])*len(imt[0])))
im.astype(float)
imt.astype(float)
# validation set
imv = im[:500]
lbv = lb[:500]
imtt = im[500:]
lbtt = lb[500:]
tag = np.zeros(len(lbv),dtype=int)
if READ:
sigma = []
mu = []
for i in range(10):
ts = np.loadtxt('sigma'+str(i)+'.out')
tm = np.loadtxt('mu'+str(i)+'.out')
sigma.append(ts)
mu.append(tm)
sigma = np.array(sigma)
mu = np.array(mu)
sigma_inv = []
for i in range(len(sigma)):
sigma_inv.append(np.linalg.inv(sigma[i]))
sigma_inv = np.array(sigma_inv)
else:
data = zip(lbtt, imtt)
datat = zip(lbv, imv)
# group data into 10 groups
data_grouped = [list(v) for l,v in groupby(sorted(data, key=lambda x:x[0]), lambda x:x[0])]
sigma, mu = cov.cov(data_grouped, 10000)
sigma = np.array(sigma)
mu = np.array(mu)
sigma_inv = []
for i in range(len(sigma)):
sigma_inv.append(np.linalg.inv(sigma[i]))
sigma_inv = np.array(sigma_inv)
#print np.linalg.slogdet(sigma[0])
pi = class_prob(lb)
cnt = 0
for imt_data in imv:
p = -sys.maxint
ind = -1
pj = class_gauss(imt_data, sigma, sigma_inv, mu)
for j in range(len(pj)):
temp = math.log(pi[j])+pj[j]
if (p < temp):
p = temp
ind = j
print lbv[cnt],':',ind
if lbv[cnt] == ind:
tag[cnt] = 1
cnt += 1
print 'hit:',sum(tag)
print 'total:', len(tag)
print 'miss rate:', 1.0-float(sum(tag))/len(tag)
def trainNetwork(self):
training_set = mnist.read(path="./mnist")
vectorized_training_set = mnist.read(path="./mnist")
self.network.train(vectorized_training_set, epochs=self.epochs, batchsize=1, eta = self.eta / 60000)
self.Status_Label.setText("Training finished!")
self.StatusLamp.setGreen()
def main():
hidden_dim = 200
batch_size = 100
print("hidden_dim: " + str(hidden_dim) + ", batch_size: " +
str(batch_size))
training_set = np.array(
list(mnist.read(dataset="training", path="/home/ywang/mnist")))
print("size of the training set: " + str(len(training_set)))
testing_set = np.array(
list(mnist.read(dataset="testing", path="/home/ywang/mnist")))
print("size of the testing set: " + str(len(testing_set)))
np.random.seed(0)
p = 1.0
backstitch_alpha = 0.3
print("backstitch alpha: " + str(backstitch_alpha))
ratio = 0.5
training_subset = training_set[np.random.binomial(1, p, len(training_set))
== 1]
print("number of the actual training examples: " +
str(len(training_subset)))
# resize the images to make the number of input features 4 times smaller
#training_images = np.stack([training_subset[i][0] for i in range(len(training_subset))])
training_images = np.stack([
ndimage.zoom(training_subset[i][0], ratio)
for i in range(len(training_subset))
])
print("image size is: " + str(training_images.shape[1]) + " by " +
str(training_images.shape[2]))
training_images = np.reshape(training_images,
[training_images.shape[0], -1])
training_labels = np.stack(
[training_subset[i][1] for i in range(len(training_subset))])
training_examples = (training_images, training_labels)
#testing_images = np.stack([testing_set[i][0] for i in range(len(testing_set))])
testing_images = np.stack([
ndimage.zoom(testing_set[i][0], ratio) for i in range(len(testing_set))
])
testing_images = np.reshape(testing_images, [testing_images.shape[0], -1])
testing_labels = np.stack(
[testing_set[i][1] for i in range(len(testing_set))])
testing_examples = (testing_images, testing_labels)
nnet = NN(num_layers=1,
input_dim=training_images.shape[1],
hidden_dim=hidden_dim,
num_classes=10,
batch_size=batch_size,
test_examples=testing_examples,
nonlin='Tanh',
update='natural',
alpha=backstitch_alpha)
nnet.Train(training_examples)
def test_numerical_gradient_checking(self):
label, image = next(mnist.read())
ninput = [pixel / 255 for row in image for pixel in row]
expected = [1 if i == label else 0 for i in range(10)]
nnet = NeuralNetwork([784, 16, 16, 10])
epsilon = 1e-5
numgrad = [np.empty(wmatrix.shape) for wmatrix in nnet.weight]
for k, wmatrix in enumerate(nnet.weight):
for i, w in np.ndenumerate(wmatrix):
wmatrix[i] = w - epsilon
nnet.feedforward(ninput)
a = nnet.get_error(expected)
wmatrix[i] = w + epsilon
nnet.feedforward(ninput)
b = nnet.get_error(expected)
numgrad[k][i] = (b - a) / 2 * epsilon
wmatrix[i] = w
error_gradient = nnet.get_error_gradient(expected)
unit = lambda v: v / norm(v) if (v != 0).any() else np.zeros(v.shape)
for k in range(len(nnet.weight)):
ag = error_gradient[k]
ng = numgrad[k]
print(f"custom = {norm(unit(ag) - unit(ng))}")
print(
f"derived from cs231 = {norm(unit(ag) * norm(ng) - ng) / max(norm(ag), norm(ng))}"
)
def learn_digits(self):
train_digits = mnist.read("training")
k_number = []
k_label = []
for i in range(5000):
k = train_digits.next()
k_label.append(k[0])
k_number.append(k[1])
y = np.array(list(k_label))
x = np.array(list(k_number))
print y[0]
print x[0]
# Now we prepare train_data and test_data.
train = x[:5000].reshape(-1,784).astype(np.float32)
# Create labels for train and test data
k = np.arange(10)
train_labels = y[:5000].astype(np.int)
# Initiate kNN, train the data, then test it with test data for k=1
knn = cv2.KNearest()
knn.train(train,train_labels)
number = self.edit_image(self.snap())
number = number.reshape(-1, 784).astype(np.float32)
nparray = np.array(number)
ret2, result2, neighbours2, dist2 = knn.find_nearest(nparray,k=5)
print result2
def main():
lb, im = mnist.read(dataset="training")
lbt, imt = mnist.read(dataset="testing")
tested = len(lbt)
tag = np.zeros(tested,dtype=int)
im = im.reshape((len(im),len(im[0])*len(im[0])))
imt = imt.reshape((len(imt),len(imt[0])*len(imt[0])))
#im = im[:100]
#imt = imt[:100]
im.astype(float)
imt.astype(float)
data = zip(lb, im)
datat = zip(lbt, imt)
# group data into 10 groups
data_grouped = [list(v) for l,v in groupby(sorted(data, key=lambda x:x[0]), lambda x:x[0])]
cov(data_grouped, 1.0)
def load_mnist(dataset_type, n_samples):
images, raw_targets = read(range(10), dataset_type)
images = images.reshape(-1, 784)[:n_samples] / 255.0
raw_targets = raw_targets[:n_samples].flatten()
targets = numpy.zeros((n_samples, 10))
targets[(range(n_samples), raw_targets)] = 1.0
return images, targets
def learn_digits(self):
train_digits = mnist.read("training")
k_number = []
k_label = []
for i in range(5000):
k = train_digits.next()
k_label.append(k[0])
k_number.append(k[1])
y = np.array(list(k_label))
x = np.array(list(k_number))
print y[0]
print x[0]
# Now we prepare train_data and test_data.
train = x[:5000].reshape(-1, 784).astype(np.float32)
# Create labels for train and test data
k = np.arange(10)
train_labels = y[:5000].astype(np.int)
# Initiate kNN, train the data, then test it with test data for k=1
knn = cv2.KNearest()
knn.train(train, train_labels)
number = self.edit_image(self.snap())
number = number.reshape(-1, 784).astype(np.float32)
nparray = np.array(number)
ret2, result2, neighbours2, dist2 = knn.find_nearest(nparray, k=5)
print result2
def run_mnist():
# FIXME: running EM on MNIST has the problem that all data collapses to one class
# This is because the likelihood for that class is slightly higher than all other.
# Probably has to do with the variance being lower for one, form k-means,
# and that being more important than closeness to mean for such high dimensional data?
# Running it with 0 iterations (i.e. on k-means) work fine, then it finds different orientations of the digits.
data_per_class = 20
training_data = list(mnist.read("training"))
dim_x, dim_y = np.shape(training_data[0][1])
ones = [d[1] for d in training_data if d[0] == 1]
fours = [d[1] for d in training_data if d[0] == 4]
fives = [d[1] for d in training_data if d[0] == 5]
ones = ones[:data_per_class]
fours = fours[:data_per_class]
fives = fives[:data_per_class]
data = np.array(ones + fours + fives).reshape((-1, dim_x * dim_y))
solver = EM(data=data, num_classes=3, num_nuisances=3)
split_data, thetas = solver.fit(max_iter=1)
for c, class_thetas in enumerate(thetas):
for n, theta in enumerate(class_thetas):
print(f"Prior: {theta.prior}, Var: {theta.variance}")
mnist.show(thetas[c][n].mean.reshape(28, 28))
def main():
lb, im = mnist.read(dataset="training")
lbt, imt = mnist.read(dataset="testing")
tested = len(lbt)
if not KDTREE:
tag = np.zeros(tested, dtype=int)
im = im.reshape((len(im), len(im[0]) * len(im[0])))
imt = imt.reshape((len(imt), len(imt[0]) * len(imt[0])))
for i in range(tested):
ind = knn_naive(imt[i], im, 10000)
if lb[ind] == lbt[i]:
tag[i] = 1
print "Good."
else:
tag[i] = 0
print "Oh no."
print "Predicted", i, ":", lb[ind]
print "Actual", i, ":", lbt[i]
print "Hit:", np.sum(tag)
print "Miss rate:", 1.0 - float(np.sum(tag)) / float(len(lbt))
else:
im = im.reshape((len(im), len(im[0]) * len(im[0])))
imt = imt.reshape((len(imt), len(imt[0]) * len(imt[0])))
if not PS:
im, lb = randomselector(im, lb, 40000)
else:
import ps
im, lb = ps.prototypeselector(im, lb, 10000)
# im,lb = prototypeselector(im,lb,1000)
print "Fitting trainning set..."
neigh = KNeighborsClassifier(n_neighbors=1)
neigh.fit(im, lb)
print "Done!"
tag = np.zeros(tested, dtype=int)
for i in range(tested):
tmp = neigh.predict(imt[i])
if tmp[0] == lbt[i]:
tag[i] = 1
print "Good."
else:
tag[i] = 0
print "Oh no."
print "Predicted", i, ":", tmp
print "Actual", i, ":", lbt[i]
print "Hit:", np.sum(tag)
print "Miss rate:", 1.0 - float(np.sum(tag)) / float(len(lbt))
def read_mnist(partial=False):
logger = logging.getLogger("mnist")
digits1 = [0, 1, 2, 3, 4]
digits2 = [5, 6, 7, 8, 9]
# Train on a subset m1 of digit1 and m2 of digit2 of the 60k training data.
if partial:
m1 = 5000
m2 = 5000
else:
m1 = 60000
m2 = 60000
# read training data
logger.info("Reading training data ...")
images, labels = mnist.read(digits1 + digits2,
dataset="training",
path=os.path.join("examples", "data"))
logger.info("done.")
def extract(images, labels):
images = images / 256.0
C1 = [k for k in xrange(len(labels)) if labels[k] in digits1]
C2 = [k for k in xrange(len(labels)) if labels[k] in digits2]
random.shuffle(C1)
random.shuffle(C2)
# Extract the random subsets together as a data matrix X (1 row per datapoint)
train = C1[:m1] + C2[:m2]
random.shuffle(train)
X = array(images[train, :])
d = array([2 * (k in digits1) - 1 for k in labels[train]])
return (X, d)
(X, d) = extract(images, labels)
logger.info("Reading test data ...")
timages, tlabels = mnist.read(digits1 + digits2,
dataset="testing",
path=os.path.join("examples", "data"))
(Xt, dt) = extract(timages, tlabels)
logger.info("done.")
return (X, d, Xt, dt)
def evaluate(path='.'):
params = train(path)
images, labels = mnist.read(range(10), 'testing', path)
c = 0
N = images.shape[0]
for i in range(N):
c += int(classify(images[i], params) == labels[i])
return c / float(N)
def evaluate(path='.'):
params = train(path)
images, labels = mnist.read(range(10), 'testing', path)
c = 0
N = images.shape[0]
for i in range(N):
c += int(classify(images[i], params) == labels[i])
return c/float(N)
def train(path='.'):
params = None
for i in range(10):
images, _ = mnist.read([i], 'training', path)
size = images[0].shape
if params is None:
params = np.zeros((10, size[0], size[1]))
params[i] = images.mean(axis=0)
return params
def test_trained(params=None, head=100, tail=100):
"Tests a network with params against first `head` and last `tail` examples"
params = params if params is not None else load_params()
nnet = NeuralNetwork(DLAYERS, params)
mnist_db = list(mnist.read())
print("[KNOWN]")
test_and_report_against(nnet, mnist_db[:head]) # Training dataset
print("[UNKNOWN]")
test_and_report_against(nnet, mnist_db[-tail:]) # Unknown dataset
def train(path='.'):
params = None
for i in range(10):
images, _ = mnist.read([i], 'training', path)
size = images[0].shape
if params is None:
params = np.zeros((10, size[0], size[1]))
params[i] = images.mean(axis=0)
return params
def read_mnist(partial=False):
logger = logging.getLogger("mnist")
digits1 = [0,1,2,3,4]
digits2 = [5,6,7,8,9]
# Train on a subset m1 of digit1 and m2 of digit2 of the 60k training data.
if partial:
m1 = 5000; m2 = 5000
else:
m1 = 60000; m2 = 60000
# read training data
logger.info("Reading training data ...")
images, labels = mnist.read(digits1 + digits2, dataset = "training",
path = os.path.join("examples", "data"))
logger.info("done.")
def extract(images, labels):
images = images / 256.0
C1 = [ k for k in xrange(len(labels)) if labels[k] in digits1 ]
C2 = [ k for k in xrange(len(labels)) if labels[k] in digits2 ]
random.shuffle(C1)
random.shuffle(C2)
# Extract the random subsets together as a data matrix X (1 row per datapoint)
train = C1[:m1] + C2[:m2]
random.shuffle(train)
X = array(images[train,:])
d = array([ 2*(k in digits1) - 1 for k in labels[train] ])
return (X,d)
(X,d) = extract(images, labels)
logger.info("Reading test data ...")
timages, tlabels = mnist.read(digits1 + digits2, dataset = "testing",
path = os.path.join("examples", "data"))
(Xt,dt) = extract(timages, tlabels)
logger.info("done.")
return (X, d, Xt, dt)
def load_mnist_all(dataset):
#load all the data
images, labels = mnist.read(digits, dataset=dataset,
path="./mnist/")
#turn cpxopt into numpy 2d array and a list
images = np.array(images)
labels = list(labels)
return normalize_mnist(images), labels
def adaboost():
X_train, y_train = read('train')
X_test, y_test = read('test')
X_train = X_train.reshape((X_train.shape[0], -1))
X_test = X_test.reshape((X_test.shape[0], -1))
bdt_discrete = AdaBoostClassifier(DecisionTreeClassifier(
max_depth=10, min_samples_split=20, min_samples_leaf=5),
n_estimators=500,
learning_rate=0.5,
algorithm='SAMME')
bdt_discrete.fit(X_train, y_train)
discrete_test_errors = []
for discrete_train_predict in bdt_discrete.staged_predict(X_test):
discrete_test_errors.append(
1. - accuracy_score(discrete_train_predict, y_test))
return bdt_discrete, discrete_test_errors
class mainWindow():
times = 1
timestart = time.clock()
to_read = [1, 8] #[0,1,2,3,4,5,6,7,8,9]
train, labels = mnist.read(to_read)
data = imresize(train[0], (400, 500))
def __init__(self):
self.root = Tkinter.Tk()
self.frame = Tkinter.Frame(self.root, width=500, height=400)
self.frame.pack()
self.canvas = Tkinter.Canvas(self.frame, width=500, height=400)
self.canvas.place(x=-2, y=-2)
self.root.after(1, self.start) # INCREASE THE 0 TO SLOW IT DOWN
self.root.mainloop()
self.shakerato = False
self.smokypixels = True
self.max_rand = 255
to_read = [1, 8] #[0,1,2,3,4,5,6,7,8,9]
self.train, self.labels = mnist.read(to_read)
self.data = imresize(self.train[0], (400, 500))
def change_img(self):
this_p = numpy.random.randint(len(self.labels))
self.train[this_p][self.labels[this_p] *
(28 / 10):(self.labels[this_p] + 1) * (28 / 10),
0:3] = 255
self.data = imresize(numpy.fliplr(self.train[this_p]), (400, 500))
def start(self):
self.im=Image.fromstring('L', (self.data.shape[1],\
self.data.shape[0]), self.data.astype('b').tostring())
self.photo = ImageTk.PhotoImage(image=self.im)
self.canvas.create_image(0, 0, image=self.photo, anchor=Tkinter.NW)
self.root.update()
self.times += 1
#### HERE REGULATES TIME BETWEEN IMAGES
if self.times % 25 == 0:
#print "%.02f FPS"%(self.times/(time.clock()-self.timestart))
self.change_img()
self.root.after(25, self.start)
#if(self.shakerato == True):
#self.data=numpy.roll(self.data,numpy.random.choice(([-1,1])),numpy.random.choice(([-1,1])))
#elif(self.smokypixels == True):
### THIS REGULATES PIXELS UPDATES
tmp_data = self.data
x, y = numpy.where(tmp_data <= 150)
tmp_data[x, y] = 130
x, y = numpy.where(tmp_data > 133) #133
for i in range(len(x) * 2):
this_p = numpy.random.randint(len(x))
self.data[x[this_p], y[this_p]] = numpy.random.randint(20)
def __init__(self):
self.root = Tkinter.Tk()
self.frame = Tkinter.Frame(self.root, width=500, height=400)
self.frame.pack()
self.canvas = Tkinter.Canvas(self.frame, width=500, height=400)
self.canvas.place(x=-2, y=-2)
self.root.after(1, self.start) # INCREASE THE 0 TO SLOW IT DOWN
self.root.mainloop()
self.shakerato = False
self.smokypixels = True
self.max_rand = 255
to_read = [1, 8] #[0,1,2,3,4,5,6,7,8,9]
self.train, self.labels = mnist.read(to_read)
self.data = imresize(self.train[0], (400, 500))
def save_mnist(dataset, dst):
label_file = open(Path(dst, "label.txt"), "w")
generator = read(dataset=dataset, path=mnist_source)
count = [0 for i in range(10)]
for label, image in generator:
filename = "{}_{:04d}.png".format(label, count[label])
label_file.write("{},{}\n".format(filename, label))
imwrite(str(Path(dst, filename)), image)
count[label] += 1
label_file.close()
def __init__(self):
self.root = Tkinter.Tk()
self.frame = Tkinter.Frame(self.root, width=500, height=400)
self.frame.pack()
self.canvas = Tkinter.Canvas(self.frame, width=500,height=400)
self.canvas.place(x=-2,y=-2)
self.root.after(1,self.start) # INCREASE THE 0 TO SLOW IT DOWN
self.root.mainloop()
self.shakerato = False
self.smokypixels = True
self.max_rand = 255
to_read = [1,8]#[0,1,2,3,4,5,6,7,8,9]
self.train , self.labels = mnist.read(to_read)
self.data = imresize(self.train[0], (400,500))
def backpropagation_main():
label, image = next(mnist.read())
ninput = [pixel / 255 for row in image for pixel in row]
expected = [1 if i == label else 0 for i in range(10)]
nnet = NeuralNetwork(DLAYERS, params=None)
# nnet = NeuralNetwork(DLAYERS, params=load_params())
for i in range(1000000000000):
guess = nnet.feedforward(ninput)
cost = nnet.get_error(expected)
print(f"[{i + 1}] cost = {cost}, guess = {guess}")
try:
nnet.backpropagate(expected)
except KeyboardInterrupt:
break
guess = nnet.feedforward(ninput)
cost = nnet.get_error(expected)
print(f"[{i + 1}] cost = {cost}")
save_params(nnet.params)
def getTrainData(count, digit):
#count is the number of training data
#digit is the specific nimber between 0 and 9 from training dataset
trainingData = np.zeros((count, 400), int)
i = 0
for img in read("training"):
if i < count:
if img[0] == digit: #This is the label
row, col = img[1].shape
im = img[1]
im = im.reshape(28, 28)
im2 = cv2.resize(im, (20, 20), interpolation=cv2.INTER_CUBIC)
im2 = im2.reshape(1, 400)
trainingData[i, :] = im2
i += 1
trainingData = (trainingData > 125).astype(int)
return trainingData
def load_mnist(dataset, digitA, digitB):
#use only the data points with labels A or B
images, labels = mnist.read(digits, dataset=dataset,
path="./mnist/")
#turn cpxopt into numpy 2d array and a list
images = np.array(images)
labels = list(labels)
new_images = []
new_labels = []
for i, label in enumerate(labels):
if label == digitA or label == digitB:
new_images.append(images[i,:])
if label == digitA:
new_labels.append(1.0)
if label == digitB:
new_labels.append(-1.0)
return normalize_mnist(np.array(new_images)), new_labels
import numpy as np
import mnist
# Load the raw MNIST
X_train, y_train = mnist.read(dataset='training')
X_test, y_test = mnist.read(dataset='testing')
# Subsample the data for more efficient code execution in this exercise
num_training = 6000
X_train = X_train[:num_training]
y_train = y_train[:num_training]
num_test = 500
X_test = X_test[:num_test]
y_test = y_test[:num_test]
# Reshape the image data into rows
# Datatype int allows you to subtract images (is otherwise uint8)
X_train = np.reshape(X_train, (X_train.shape[0], -1)).astype('int')
X_test = np.reshape(X_test, (X_test.shape[0], -1)).astype('int')
# As a sanity check, we print out the size of the training and test data.
print('Training data shape: ', X_train.shape)
print('Training labels shape: ', y_train.shape)
print('Test data shape: ', X_test.shape)
print('Test labels shape: ', y_test.shape)
###############################################################################
# #
# Implement the k Nearest Neighbors algorithm here #
# Do we have a fancy GPU?
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
# Parameters
num_epochs = 100
num_classes = 10 # i.e. the digits
batch_size = 300 # How many samples per batch use
learning_rate = 0.001
# Load MNIST data into
X_train = []
X_test = []
Y_train = []
Y_test = []
for element in m.read("training"):
Y_train.append(element[0])
X_train.append(element[1])
for element in m.read("testing"):
Y_test.append(element[0])
X_test.append(element[1])
X_train = np.expand_dims(np.asarray(X_train, dtype=np.float32), axis=1)
X_test = np.expand_dims(np.asarray(X_test, dtype=np.float32), axis=1)
Y_train = np.asarray(Y_train, dtype=np.int64)
Y_test = np.asarray(Y_test, dtype=np.int64)
train_set = torch.utils.data.TensorDataset(torch.from_numpy(X_train),
torch.from_numpy(Y_train))
test_set = torch.utils.data.TensorDataset(torch.from_numpy(X_test),
from mnist import read, show, normalize
import numpy as np
import matplotlib.pyplot as plt
train = list(read('train'))
data = train[0]
label = data[0]
pixels = data[1:]
pixels = np.array(pixels, dtype='uint8')
pixels = pixels.reshape((28, 28))
plt.title('Example of MNIST pattern')
plt.imshow(pixels, cmap='gray')
plt.show()
train_obj = softmaxObjective(train_x, train_y, theta)
print('Iteration {}: Training set objective function: {} \r'.format(
ct, train_obj))
theta += -(alpha * gradient)
assert theta.shape == (n, 10)
ct += 1
print ""
print("Training set accuracy: {0:0.2f}%".format(
100 * softmax_classifier_acuracy(theta, train_x, train_y)))
print("Testing set accuracy: {0:0.2f}%".format(
100 * softmax_classifier_acuracy(theta, test_x, test_y)))
cwd = os.getcwd()
# "training" or "testing"
train_images = read("training", cwd)
n = 28 * 28
m = 1000 #No of training sets, max: 60000
train_y = np.empty([m, 1])
train_x = np.empty([m, n])
i = 0
for image in train_images:
if i >= m: break
train_x[i, :] = np.reshape(image[1], (1, n))
train_y[i] = image[0]
i += 1
train_x = train_x[:i, :]
train_y = train_y[:i]
mean_x = np.mean(train_x, axis=0)
std_x = np.std(train_x, axis=0) + 0.1
"""
Implements a Neural Network
"""
from vectorflux import VectorFlux
from mnist import read, show, normalize
from vectorflux.layers import Dense
from vectorflux.layers.Dropout import Dropout
train = list(read('train'))
test = list(read('test'))
print("Train size: {}".format(len(train)))
print("Test size: {}".format(len(test)))
# Normalization for values
test_x, test_y = normalize(test)
train_x, train_y = normalize(train)
vf = VectorFlux()
vf.add(Dense(800, activation='sigmoid', input_shape=784, optimizer='Momentum'))
vf.add(Dropout(0.5, input_shape=800))
vf.add(Dense(800, activation='sigmoid', input_shape=800, optimizer='ADAM'))
vf.add(Dense(10, activation='sigmoid', input_shape=800))
vf.train(x_train=train_x,
y_train=train_y,
x_test=test_x,
y_test=test_y,
epochs=100000,
def test(self):
images, labels = mnist.read('testing')
test(images, labels)
def main():
lbt, imt = mnist.read(dataset="testing")
print lbt[4]
mnist.show(imt[4])
class GANeuralNetwork(Subject, NeuralNetwork):
__mnist_db = list(mnist.read())
_genome: List[float]
_fitness: float
def __init__(self, params):
"""
Precondition: use set_layers_description() before any instanciation
so dlayers is initialized
"""
super().__init__(GANeuralNetwork.dlayers, params)
self._genome = params
self._fitness = None
@classmethod
def create_random(cls):
return GANeuralNetwork(cls.get_random_params())
@property
def genome(self) -> List[float]:
return self._genome
@property
def fitness(self) -> float:
return self.batch_cost() if not self._fitness else self._fitness
def batch_cost(self, batch_size=10, random_samples=False):
"Runs a random minibatch and returns average network cost"
costs = [None] * batch_size
db = (
sample(GANeuralNetwork.__mnist_db, batch_size) if random_samples
else GANeuralNetwork.__mnist_db[:batch_size]
)
for i, (label, image) in enumerate(db): # TODO: parallelize runs
# Run network
ninput = [pixel / 255 for row in image for pixel in row] # Normalized
guess = self(self.genome, ninput)
# Cost calculation
expected = [1 if i == label else 0 for i in range(10)]
costs[i] = sum((g - e)**2 for g, e in zip(guess, expected))
cost = mean(costs)
self._fitness = -cost
# print(f"Average cost of {cost} after {batch_size} runs")
return self._fitness
# TODO: Think more about this and make it
# Maybe a urand in [c +- d] range with c = (min + max) / 2, d = max - min
@staticmethod
def mutate(gen):
return gen + randn()
@classmethod
def set_layers_description(cls, dlayers):
"""
Override of NeuralNetwork method that makes it static
dlayers will be used as a static attribute of GANeuralNetwork class
"""
cls.dlayers = dlayers
@classmethod
def get_random_params(cls):
return super().get_random_params(cls)
def train(self):
images, labels = mnist.read('training')
train(images, labels)
@author: Bruce """ import os import mnist, svmcmpl, cvxopt, random digits1 = [ 0 ] digits2 = [ 1 ] m1 = 4000; m2 = 4000 # read training data dpath = os.getcwd()+"\\data\\mnist" print( dpath) images, labels = mnist.read(digits1 + digits2, dataset = "training", path = dpath ) images = images / 256. C1 = [ k for k in range(len(labels)) if labels[k] in digits1 ] C2 = [ k for k in range(len(labels)) if labels[k] in digits2 ] random.seed() random.shuffle(C1) random.shuffle(C2) train = C1[:m1] + C2[:m2] random.shuffle(train) X = images[train,:] d = cvxopt.matrix([ 2*(k in digits1) - 1 for k in labels[train] ]) gamma = 4.0
import sys
import numpy as np
import scipy.misc
import imageio
sys.path.append("modules")
import mnist
sys.path.append("extension")
import pyceptron
training_set = mnist.read(dataset="training", path="mnist")
testing_set = mnist.read(dataset="testing", path="mnist")
architecture = [784, 300, 10]
activation = "ReLU"
softmax = 1
network = pyceptron.Network(architecture,
activation=activation,
softmax=softmax)
network.load_state("states/784_300_10_sgd_ReLU_softmax_229.state")
bad_lbl = list()
bad_img = list()
print(len(testing_set))
# Calculate the training error rate
for i in range(len(testing_set)):
sample = testing_set[i]
prediction = network.predict(sample[0])
prediction = np.argmax(prediction)
expectation = np.argmax(sample[1])
if not prediction == expectation:
import mnist
import Image
import os
from pylab import imread, imshow, gray, mean
from numpy import array
from cvxopt import matrix
from scipy.misc import imsave
alphabets = ['a', 'b', 'c', 'A']
# read all the training data and labels
for i in alphabets:
print 'Reading char=' + str(i) + ', ord=' + str(ord(i))
images, labels = mnist.read([ord(i)])
for j in range(0, images.size[0]):
img_matrix_1d = images[j,:]
img_matrix = matrix(img_matrix_1d, (28,28)).trans()
img_array = array(img_matrix)
savePath = './raw/training/'+str(ord(i))+'/'
if not os.path.exists(savePath):
os.makedirs(savePath)
imsave(savePath + str(j)+'.png', img_array)
from scipy.misc import imsave from sklearn import svm, metrics from sklearn.cross_validation import LeaveOneOut from sklearn.externals import joblib from sklearn.multiclass import OneVsRestClassifier import prepare_data import scipy from numpy import vstack from numpy import append import os # train the digits 0 alphabets =[] #['a','b','c','d','e','f','g','h','i','j','k','l','m','n','o','p','q','r','s','t','u','v','w','x','y','z'] #'a', 'd', 'e', 'c', 'u', 's', 'o', 'f', 't'] alphabets_ord = map(ord, alphabets) ignoreList = ['0','1', '2', '3', '4', '5', '6', '7', '8', '9'] ignoreList = map(ord, ignoreList) images, labels = mnist.read(alphabets_ord, ignoreList = ignoreList) images = array(images) labels = array(labels).reshape(1, len(labels))[0] fonts = ['nothingyoucoulddo', 'comicsans', 'dakota', 'showhands', 'danielbd', 'danielbk', 'dandelion', 'daniel' ] for font in fonts: font_images, font_labels = mnist.read(alphabets_ord, './../../data/', font + '_img.idx', font + '_label.idx', ignoreList=ignoreList) font_images = array(font_images) font_labels = array(font_labels).reshape(1, len(font_labels))[0] images = vstack((images, font_images)) labels = append(labels, font_labels) ''' images, labels = mnist.read(alphabets_ord, './../../data/', 'comicsans_img.idx', 'comicsans_label.idx')
def __init__(self, config, phase='train'):
self.config = config
self.imgs, self.lbls = mnist.read(dataset=phase,
path=self.config.mnist_dir)
import os
import mnist
from cv2 import imwrite
src = "raw_mnist"
def initialize_folders():
for i in range(10):
folder_name = os.path.join(src, str(i))
os.makedirs(folder_name, exist_ok=True)
if __name__ == "__main__":
initialize_folders()
gen = mnist.read()
counter = 0
while True:
label, img = next(gen)
dir_name = os.path.join(src, str(label))
filename = os.path.join(src, str(label), "{}_{:04d}.png".format(label, len(os.listdir(dir_name))))
imwrite(filename, img)
counter+=1
if counter%1000 == 0:
print(counter)
def main():
f = 0.05
th = 6.27344414456
#f = 0.1
#th = 9.25216525759
#f = 0.15
#th = 12.6067380436
#f = 0.2
#th = 14.6211445211
lb, im = mnist.read(dataset="training")
lbt, imt = mnist.read(dataset="testing")
tested = len(lbt)
tag = np.zeros(tested,dtype=int)
im = im.reshape((len(im),len(im[0])*len(im[0])))
imt = imt.reshape((len(imt),len(imt[0])*len(imt[0])))
if READ:
sigma = []
mu = []
for i in range(10):
ts = np.loadtxt('sigma'+str(i)+'.out')
tm = np.loadtxt('mu'+str(i)+'.out')
sigma.append(ts)
mu.append(tm)
sigma = np.array(sigma)
mu = np.array(mu)
sigma_inv = []
for i in range(len(sigma)):
sigma_inv.append(np.linalg.inv(sigma[i]))
sigma_inv = np.array(sigma_inv)
else:
data = zip(lb, im)
datat = zip(lbt, imt)
# group data into 10 groups
data_grouped = [list(v) for l,v in groupby(sorted(data, key=lambda x:x[0]), lambda x:x[0])]
sigma, mu = cov.cov(data_grouped, 10000)
sigma = np.array(sigma)
mu = np.array(mu)
sigma_inv = []
for i in range(len(sigma)):
sigma_inv.append(np.linalg.inv(sigma[i]))
sigma_inv = np.array(sigma_inv)
#print np.linalg.slogdet(sigma[0])
pi = class_prob(lb)
cnt = 0
ab = 0
for imt_data in imt:
p = -sys.maxint
p2 = -sys.maxint
ind = -1
ind2 = -1
pj = class_gauss(imt_data, sigma, sigma_inv, mu)
for j in range(len(pj)):
temp = math.log(pi[j])+pj[j]
if (p < temp):
p = temp
ind = j
elif (p2 < temp):
p2 = temp
ind2 = j
if (p-p2) < th:
ab += 1
print lbt[cnt],':',ind, ':', p, '(abstain!)'
else:
print lbt[cnt],':',ind, ':', p
if lbt[cnt] == ind:
tag[cnt] = 1
cnt += 1
print 'hit:',sum(tag)
print 'total:', len(tag)
print 'abstain:', ab
print 'miss rate:', 1.0-float(sum(tag))/(len(tag)-ab)
import argparse
import mnist
import naive_bayes as nb
import numpy as np
parser = argparse.ArgumentParser(description='It is a program for ML HW#2.')
parser.add_argument('train_img_path', help='file path of train img', type=str)
parser.add_argument('train_lbl_path', help='file path of train lbl', type=str)
parser.add_argument('test_img_path', help='file path of test img', type=str)
parser.add_argument('test_lbl_path', help='file path of test lbl', type=str)
parser.add_argument('mode', help='toggle option', type=int)
args = parser.parse_args()
print('train_img_path: {}'.format(args.train_img_path))
print('train_lbl_path: {}'.format(args.train_lbl_path))
print('test_img_path: {}'.format(args.test_img_path))
print('test_img_path: {}'.format(args.test_lbl_path))
train_img, train_lbl = mnist.read(args.train_img_path, args.train_lbl_path)
test_img, test_lbl = mnist.read(args.test_img_path, args.test_lbl_path)
print(train_img.shape)
nb.classify(train_img, train_lbl, test_img, test_lbl, args.mode)
# Dense data
#y, x = [1,-1], [[1,0,1], [-1,0,-1]]
# Sparse data
y, x = [1,-1], [{1:1, 3:1}, {1:-1,3:-1}]
prob = svm_problem(y, x)
param = svm_parameter('-t 0 -c 4 -b 1')
m = svm_train(prob, param)
'''
SUBSAMPLE = 10000
digits = [0,1,2,3,4,5,6,7,8,9]
DATAPATH = "./data/mnist/"
train_images, train_labels = mnist.read(digits, dataset='training', path=DATAPATH)
x = np.array(train_images).tolist() #svm requires a list
y = np.array(train_labels).T.astype(float).tolist()[0]
pprint(len(x))
x = x[:SUBSAMPLE]
y = y[:SUBSAMPLE]
#pprint(x)
#pprint(y)
#pprint(x[0])
prob = svm_problem(y, x)
return fin
if __name__ == '__main__':
config = Config()
n_gpu = pytorch_utils.setgpu('6')
net = ResNet_CAM()
checkpoint = torch.load(root_path +
'checkpoints/020.ckpt') # must before cuda
net.load_state_dict(checkpoint['state_dict'])
net = net.cuda()
cudnn.benchmark = True
net = DataParallel(net)
net.eval()
imgs, lbls = mnist.read(dataset='test', path=config.mnist_dir)
idcs = np.random.randint(0, len(lbls), size=(3, ))
img_pad = np.zeros([3, config.img_size, config.img_size], dtype=np.float32)
classes = []
for idx in idcs:
img = imgs[idx].astype(np.float32)
ih, iw = img.shape
img = (img / 255. - config.mnist_mean) / config.mnist_std
x, y = np.random.randint(0,
config.img_size - ih, (2, ),
dtype=np.int16)
for k in range(3):
img_pad[k, y:y + ih, x:x + iw] = img
classes.append(lbls[idx])
data = np.expand_dims(img_pad, 0).copy()
data = torch.from_numpy(data)
import mnist
import numpy as np
import matplotlib.pyplot as plt
plt.figure()
images = mnist.read()
for i in range(16):
plt.subplot(4,4,i+1)
im = images.next()[1]
fig = plt.imshow(1 - im/128, cmap='Greys')
fig.axes.get_xaxis().set_visible(False)
fig.axes.get_yaxis().set_visible(False)
plt.savefig('ex.png', bbox_inches='tight', pad_inches=0, dpi=300)
#fig.show()
import mnist, Image, sys
from pylab import imread, imshow, gray, mean
from numpy import array
from cvxopt import matrix
from pprint import pprint
from scipy.misc import imsave
from sklearn.metrics import classification_report, precision_score, recall_score
from sklearn.cross_validation import LeaveOneOut, StratifiedKFold
from sklearn.grid_search import GridSearchCV
from sklearn.svm import SVC, LinearSVC
# train the digits 0
alphabets = [] # ['a', 'b', 'c', 'd', 'e', 'A', 'B', 'C', 'D', 'E']
alphabets_ord = map(ord, alphabets)
images, labels = mnist.read(alphabets_ord)
images = array(images)
labels = array(labels).reshape(1, len(labels))[0]
n_samples = len(images)
# read the training data and labelsi
sys.stdout.write('Reading in training data and labels')
X = images
y = labels
sys.stdout.write(' ... Done!\n')
# split the data into two equal parts respecting label proprtions
train, test = iter(StratifiedKFold(y, 2)).next()
#################################################
# set the tuning parameters
from keras.models import Sequential
from keras.layers import Dense
import keras
import numpy as np
import mnist
import time
inp = 784
img,lbl = mnist.read(dataset = "training", path = "")
img = img[:1000]
lbl = lbl[:1000]
x=[]
for u in img:
v = []
for l in u:
v.extend(l/255.0)
x.append(v)
x = np.array(x)
y = []
for a in lbl:
v = [0.0]*10
v[a] = 1.0
y.append(v)
y = np.array(y)
model = Sequential()
model.add(Dense(100, input_dim=inp, activation='sigmoid'))
import mnist, Image, sys
from pylab import imread, imshow, gray, mean
from numpy import array
from cvxopt import matrix
from scipy.misc import imsave
from sklearn import svm, metrics
# train the digits 0
digit = [1,2,3,4,5]
# read the training data and labelsi
sys.stdout.write('Reading in training data and labels')
train_images, train_labels = mnist.read(digit, "training", "./../../data")
sys.stdout.write(' ... Done!\n')
# read the test data and labels
sys.stdout.write('Reading in testing data and labels')
test_images, test_labels = mnist.read(digit, "testing", "./../../data")
sys.stdout.write(' ... Done!\n')
# convert to arrays
x_train = array(train_images)
y_train = array(train_labels).reshape(1, len(train_labels))[0]
x_test = array(test_images)
y_test = array(test_labels).reshape(1, len(test_labels))[0]
# create classifier
classifier = svm.LinearSVC()
# train the classifier
import mnist
import Image
from pylab import imread, imshow, gray, mean
from numpy import array
from cvxopt import matrix
from scipy.misc import imsave
# read all the training data and labels
images, labels = mnist.read([0])
# get the first row of pixel dataq
img0_matrix_1d = images[0,:]
img0_matrix = matrix(img0_matrix_1d, (28,28))
img0_array = array(img0_matrix)
# save the image
imsave('./raw/mnist_training_digit_4_data_0.png', img0_array)
